Limitations of Foundation Models in Energy Materials Simulations: A Case Study in Polyanion Sodium Cathode Materials

Recently, universal machine learning interatomic potentials (MLIPs) are being increasingly adopted to study various physical properties of promising new materials. To enhance the energy density and ionic conductivity of sodium‐ion batteries (SIBs), polyanion‐type cathode materials with transition metal ion doping are being explored, for which MLIPs can be used. Even with broad spectrum training data, universal MLIPs still suffer from inaccurate predictions in out‐of‐distribution systems. We developed an MLIP based on an update to the popular PaiNN(charge‐PaiNN, cPaiNN) architecture to include a description of atomic charges that can work with training datasets with partial labeling of atomic charges. By comparing a cPaiNN model specifically trained on polyanion sodium cathode materials against three universal MLIPs, CHGNet, MACE‐MP‐0, and M3GNet, we demonstrate the necessity of system‐specific datasets to improve model accuracy. We further demonstrate that fine‐tuning universal MLIPs on system‐specific datasets improves performance, achieving accuracy comparable to or better than training from scratch. Nonetheless, cPaiNN generally outperforms all foundation models in both accuracy and computational speed on the system specific test dataset. Furthermore, we check the importance of atomic charge predictions for accurately predicting the physical properties of cathode materials.